search

janosh / matbench-discovery

An evaluation framework for machine learning models simulating high-throughput materials discovery.
https://matbench-discovery.materialsproject.org
MIT License
111 stars 18 forks source link

Add MatterSim predictions #104

Closed yanghan-microsoft closed 5 months ago

yanghan-microsoft commented 5 months ago

Dear developers,

We are submitting a pull request with the intention of adding the predictions of our recently developed model, MatterSim, to the leaderboard. This submission includes the requisite compressed data for evaluation purposes, along with our internally tested performance metrics for your reference.

Should there be any inconsistencies between the metrics we have provided and those obtained from your evaluations, we would appreciate being notified. We wish to express our sincere gratitude for your diligent efforts in developing this leaderboad, which has proven to be incredibly beneficial to our entire community.

Best regards,

Han Yang Materials team Microsoft Research AI for Science

ZLMSFT commented 5 months ago

@yanghan-microsoft Han, it is Okay to upload the test.py (correct me if wrong) to the repo. As long as we do not leak the weights for now, it is fine. No worries. Let’s make the model up on the board asap.

@janosh Hi Janosh, thanks for the help and understanding. We were basically following the practice of the GNOME model. But since you asked, it is indeed good to give more information that are directly relevant to the paper in the code here.